root/net/core/xdp.c

DEFINITIONS

1. xdp_mem_id_hashfn
2. xdp_mem_id_cmp
3. __xdp_mem_allocator_rcu_free
4. mem_xa_remove
5. mem_allocator_disconnect
6. mem_id_disconnect
7. xdp_rxq_info_unreg_mem_model
8. xdp_rxq_info_unreg
9. xdp_rxq_info_init
10. xdp_rxq_info_reg
11. xdp_rxq_info_unused
12. xdp_rxq_info_is_reg
13. __mem_id_init_hash_table
14. __mem_id_cyclic_get
15. __is_supported_mem_type
16. xdp_rxq_info_reg_mem_model
17. __xdp_return
18. xdp_return_frame
19. xdp_return_frame_rx_napi
20. xdp_return_buff
21. __xdp_release_frame
22. xdp_attachment_query
23. xdp_attachment_flags_ok
24. xdp_attachment_setup
25. xdp_convert_zc_to_xdp_frame

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /* net/core/xdp.c
   3  *
   4  * Copyright (c) 2017 Jesper Dangaard Brouer, Red Hat Inc.
   5  */
   6 #include <linux/bpf.h>
   7 #include <linux/filter.h>
   8 #include <linux/types.h>
   9 #include <linux/mm.h>
  10 #include <linux/netdevice.h>
  11 #include <linux/slab.h>
  12 #include <linux/idr.h>
  13 #include <linux/rhashtable.h>
  14 #include <net/page_pool.h>
  15 
  16 #include <net/xdp.h>
  17 #include <net/xdp_priv.h> /* struct xdp_mem_allocator */
  18 #include <trace/events/xdp.h>
  19 
  20 #define REG_STATE_NEW           0x0
  21 #define REG_STATE_REGISTERED    0x1
  22 #define REG_STATE_UNREGISTERED  0x2
  23 #define REG_STATE_UNUSED        0x3
  24 
  25 static DEFINE_IDA(mem_id_pool);
  26 static DEFINE_MUTEX(mem_id_lock);
  27 #define MEM_ID_MAX 0xFFFE
  28 #define MEM_ID_MIN 1
  29 static int mem_id_next = MEM_ID_MIN;
  30 
  31 static bool mem_id_init; /* false */
  32 static struct rhashtable *mem_id_ht;
  33 
  34 static u32 xdp_mem_id_hashfn(const void *data, u32 len, u32 seed)
  35 {
  36         const u32 *k = data;
  37         const u32 key = *k;
  38 
  39         BUILD_BUG_ON(FIELD_SIZEOF(struct xdp_mem_allocator, mem.id)
  40                      != sizeof(u32));
  41 
  42         /* Use cyclic increasing ID as direct hash key */
  43         return key;
  44 }
  45 
  46 static int xdp_mem_id_cmp(struct rhashtable_compare_arg *arg,
  47                           const void *ptr)
  48 {
  49         const struct xdp_mem_allocator *xa = ptr;
  50         u32 mem_id = *(u32 *)arg->key;
  51 
  52         return xa->mem.id != mem_id;
  53 }
  54 
  55 static const struct rhashtable_params mem_id_rht_params = {
  56         .nelem_hint = 64,
  57         .head_offset = offsetof(struct xdp_mem_allocator, node),
  58         .key_offset  = offsetof(struct xdp_mem_allocator, mem.id),
  59         .key_len = FIELD_SIZEOF(struct xdp_mem_allocator, mem.id),
  60         .max_size = MEM_ID_MAX,
  61         .min_size = 8,
  62         .automatic_shrinking = true,
  63         .hashfn    = xdp_mem_id_hashfn,
  64         .obj_cmpfn = xdp_mem_id_cmp,
  65 };
  66 
  67 static void __xdp_mem_allocator_rcu_free(struct rcu_head *rcu)
  68 {
  69         struct xdp_mem_allocator *xa;
  70 
  71         xa = container_of(rcu, struct xdp_mem_allocator, rcu);
  72 
  73         /* Allow this ID to be reused */
  74         ida_simple_remove(&mem_id_pool, xa->mem.id);
  75 
  76         /* Poison memory */
  77         xa->mem.id = 0xFFFF;
  78         xa->mem.type = 0xF0F0;
  79         xa->allocator = (void *)0xDEAD9001;
  80 
  81         kfree(xa);
  82 }
  83 
  84 static void mem_xa_remove(struct xdp_mem_allocator *xa)
  85 {
  86         trace_mem_disconnect(xa);
  87 
  88         if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
  89                 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
  90 }
  91 
  92 static void mem_allocator_disconnect(void *allocator)
  93 {
  94         struct xdp_mem_allocator *xa;
  95         struct rhashtable_iter iter;
  96 
  97         mutex_lock(&mem_id_lock);
  98 
  99         rhashtable_walk_enter(mem_id_ht, &iter);
 100         do {
 101                 rhashtable_walk_start(&iter);
 102 
 103                 while ((xa = rhashtable_walk_next(&iter)) && !IS_ERR(xa)) {
 104                         if (xa->allocator == allocator)
 105                                 mem_xa_remove(xa);
 106                 }
 107 
 108                 rhashtable_walk_stop(&iter);
 109 
 110         } while (xa == ERR_PTR(-EAGAIN));
 111         rhashtable_walk_exit(&iter);
 112 
 113         mutex_unlock(&mem_id_lock);
 114 }
 115 
 116 static void mem_id_disconnect(int id)
 117 {
 118         struct xdp_mem_allocator *xa;
 119 
 120         mutex_lock(&mem_id_lock);
 121 
 122         xa = rhashtable_lookup_fast(mem_id_ht, &id, mem_id_rht_params);
 123         if (!xa) {
 124                 mutex_unlock(&mem_id_lock);
 125                 WARN(1, "Request remove non-existing id(%d), driver bug?", id);
 126                 return;
 127         }
 128 
 129         trace_mem_disconnect(xa);
 130 
 131         if (!rhashtable_remove_fast(mem_id_ht, &xa->node, mem_id_rht_params))
 132                 call_rcu(&xa->rcu, __xdp_mem_allocator_rcu_free);
 133 
 134         mutex_unlock(&mem_id_lock);
 135 }
 136 
 137 void xdp_rxq_info_unreg_mem_model(struct xdp_rxq_info *xdp_rxq)
 138 {
 139         struct xdp_mem_allocator *xa;
 140         int id = xdp_rxq->mem.id;
 141 
 142         if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
 143                 WARN(1, "Missing register, driver bug");
 144                 return;
 145         }
 146 
 147         if (id == 0)
 148                 return;
 149 
 150         if (xdp_rxq->mem.type == MEM_TYPE_ZERO_COPY)
 151                 return mem_id_disconnect(id);
 152 
 153         if (xdp_rxq->mem.type == MEM_TYPE_PAGE_POOL) {
 154                 rcu_read_lock();
 155                 xa = rhashtable_lookup(mem_id_ht, &id, mem_id_rht_params);
 156                 page_pool_destroy(xa->page_pool);
 157                 rcu_read_unlock();
 158         }
 159 }
 160 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg_mem_model);
 161 
 162 void xdp_rxq_info_unreg(struct xdp_rxq_info *xdp_rxq)
 163 {
 164         /* Simplify driver cleanup code paths, allow unreg "unused" */
 165         if (xdp_rxq->reg_state == REG_STATE_UNUSED)
 166                 return;
 167 
 168         WARN(!(xdp_rxq->reg_state == REG_STATE_REGISTERED), "Driver BUG");
 169 
 170         xdp_rxq_info_unreg_mem_model(xdp_rxq);
 171 
 172         xdp_rxq->reg_state = REG_STATE_UNREGISTERED;
 173         xdp_rxq->dev = NULL;
 174 
 175         /* Reset mem info to defaults */
 176         xdp_rxq->mem.id = 0;
 177         xdp_rxq->mem.type = 0;
 178 }
 179 EXPORT_SYMBOL_GPL(xdp_rxq_info_unreg);
 180 
 181 static void xdp_rxq_info_init(struct xdp_rxq_info *xdp_rxq)
 182 {
 183         memset(xdp_rxq, 0, sizeof(*xdp_rxq));
 184 }
 185 
 186 /* Returns 0 on success, negative on failure */
 187 int xdp_rxq_info_reg(struct xdp_rxq_info *xdp_rxq,
 188                      struct net_device *dev, u32 queue_index)
 189 {
 190         if (xdp_rxq->reg_state == REG_STATE_UNUSED) {
 191                 WARN(1, "Driver promised not to register this");
 192                 return -EINVAL;
 193         }
 194 
 195         if (xdp_rxq->reg_state == REG_STATE_REGISTERED) {
 196                 WARN(1, "Missing unregister, handled but fix driver");
 197                 xdp_rxq_info_unreg(xdp_rxq);
 198         }
 199 
 200         if (!dev) {
 201                 WARN(1, "Missing net_device from driver");
 202                 return -ENODEV;
 203         }
 204 
 205         /* State either UNREGISTERED or NEW */
 206         xdp_rxq_info_init(xdp_rxq);
 207         xdp_rxq->dev = dev;
 208         xdp_rxq->queue_index = queue_index;
 209 
 210         xdp_rxq->reg_state = REG_STATE_REGISTERED;
 211         return 0;
 212 }
 213 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg);
 214 
 215 void xdp_rxq_info_unused(struct xdp_rxq_info *xdp_rxq)
 216 {
 217         xdp_rxq->reg_state = REG_STATE_UNUSED;
 218 }
 219 EXPORT_SYMBOL_GPL(xdp_rxq_info_unused);
 220 
 221 bool xdp_rxq_info_is_reg(struct xdp_rxq_info *xdp_rxq)
 222 {
 223         return (xdp_rxq->reg_state == REG_STATE_REGISTERED);
 224 }
 225 EXPORT_SYMBOL_GPL(xdp_rxq_info_is_reg);
 226 
 227 static int __mem_id_init_hash_table(void)
 228 {
 229         struct rhashtable *rht;
 230         int ret;
 231 
 232         if (unlikely(mem_id_init))
 233                 return 0;
 234 
 235         rht = kzalloc(sizeof(*rht), GFP_KERNEL);
 236         if (!rht)
 237                 return -ENOMEM;
 238 
 239         ret = rhashtable_init(rht, &mem_id_rht_params);
 240         if (ret < 0) {
 241                 kfree(rht);
 242                 return ret;
 243         }
 244         mem_id_ht = rht;
 245         smp_mb(); /* mutex lock should provide enough pairing */
 246         mem_id_init = true;
 247 
 248         return 0;
 249 }
 250 
 251 /* Allocate a cyclic ID that maps to allocator pointer.
 252  * See: https://www.kernel.org/doc/html/latest/core-api/idr.html
 253  *
 254  * Caller must lock mem_id_lock.
 255  */
 256 static int __mem_id_cyclic_get(gfp_t gfp)
 257 {
 258         int retries = 1;
 259         int id;
 260 
 261 again:
 262         id = ida_simple_get(&mem_id_pool, mem_id_next, MEM_ID_MAX, gfp);
 263         if (id < 0) {
 264                 if (id == -ENOSPC) {
 265                         /* Cyclic allocator, reset next id */
 266                         if (retries--) {
 267                                 mem_id_next = MEM_ID_MIN;
 268                                 goto again;
 269                         }
 270                 }
 271                 return id; /* errno */
 272         }
 273         mem_id_next = id + 1;
 274 
 275         return id;
 276 }
 277 
 278 static bool __is_supported_mem_type(enum xdp_mem_type type)
 279 {
 280         if (type == MEM_TYPE_PAGE_POOL)
 281                 return is_page_pool_compiled_in();
 282 
 283         if (type >= MEM_TYPE_MAX)
 284                 return false;
 285 
 286         return true;
 287 }
 288 
 289 int xdp_rxq_info_reg_mem_model(struct xdp_rxq_info *xdp_rxq,
 290                                enum xdp_mem_type type, void *allocator)
 291 {
 292         struct xdp_mem_allocator *xdp_alloc;
 293         gfp_t gfp = GFP_KERNEL;
 294         int id, errno, ret;
 295         void *ptr;
 296 
 297         if (xdp_rxq->reg_state != REG_STATE_REGISTERED) {
 298                 WARN(1, "Missing register, driver bug");
 299                 return -EFAULT;
 300         }
 301 
 302         if (!__is_supported_mem_type(type))
 303                 return -EOPNOTSUPP;
 304 
 305         xdp_rxq->mem.type = type;
 306 
 307         if (!allocator) {
 308                 if (type == MEM_TYPE_PAGE_POOL || type == MEM_TYPE_ZERO_COPY)
 309                         return -EINVAL; /* Setup time check page_pool req */
 310                 return 0;
 311         }
 312 
 313         /* Delay init of rhashtable to save memory if feature isn't used */
 314         if (!mem_id_init) {
 315                 mutex_lock(&mem_id_lock);
 316                 ret = __mem_id_init_hash_table();
 317                 mutex_unlock(&mem_id_lock);
 318                 if (ret < 0) {
 319                         WARN_ON(1);
 320                         return ret;
 321                 }
 322         }
 323 
 324         xdp_alloc = kzalloc(sizeof(*xdp_alloc), gfp);
 325         if (!xdp_alloc)
 326                 return -ENOMEM;
 327 
 328         mutex_lock(&mem_id_lock);
 329         id = __mem_id_cyclic_get(gfp);
 330         if (id < 0) {
 331                 errno = id;
 332                 goto err;
 333         }
 334         xdp_rxq->mem.id = id;
 335         xdp_alloc->mem  = xdp_rxq->mem;
 336         xdp_alloc->allocator = allocator;
 337 
 338         /* Insert allocator into ID lookup table */
 339         ptr = rhashtable_insert_slow(mem_id_ht, &id, &xdp_alloc->node);
 340         if (IS_ERR(ptr)) {
 341                 ida_simple_remove(&mem_id_pool, xdp_rxq->mem.id);
 342                 xdp_rxq->mem.id = 0;
 343                 errno = PTR_ERR(ptr);
 344                 goto err;
 345         }
 346 
 347         if (type == MEM_TYPE_PAGE_POOL)
 348                 page_pool_use_xdp_mem(allocator, mem_allocator_disconnect);
 349 
 350         mutex_unlock(&mem_id_lock);
 351 
 352         trace_mem_connect(xdp_alloc, xdp_rxq);
 353         return 0;
 354 err:
 355         mutex_unlock(&mem_id_lock);
 356         kfree(xdp_alloc);
 357         return errno;
 358 }
 359 EXPORT_SYMBOL_GPL(xdp_rxq_info_reg_mem_model);
 360 
 361 /* XDP RX runs under NAPI protection, and in different delivery error
 362  * scenarios (e.g. queue full), it is possible to return the xdp_frame
 363  * while still leveraging this protection.  The @napi_direct boolian
 364  * is used for those calls sites.  Thus, allowing for faster recycling
 365  * of xdp_frames/pages in those cases.
 366  */
 367 static void __xdp_return(void *data, struct xdp_mem_info *mem, bool napi_direct,
 368                          unsigned long handle)
 369 {
 370         struct xdp_mem_allocator *xa;
 371         struct page *page;
 372 
 373         switch (mem->type) {
 374         case MEM_TYPE_PAGE_POOL:
 375                 rcu_read_lock();
 376                 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
 377                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 378                 page = virt_to_head_page(data);
 379                 napi_direct &= !xdp_return_frame_no_direct();
 380                 page_pool_put_page(xa->page_pool, page, napi_direct);
 381                 rcu_read_unlock();
 382                 break;
 383         case MEM_TYPE_PAGE_SHARED:
 384                 page_frag_free(data);
 385                 break;
 386         case MEM_TYPE_PAGE_ORDER0:
 387                 page = virt_to_page(data); /* Assumes order0 page*/
 388                 put_page(page);
 389                 break;
 390         case MEM_TYPE_ZERO_COPY:
 391                 /* NB! Only valid from an xdp_buff! */
 392                 rcu_read_lock();
 393                 /* mem->id is valid, checked in xdp_rxq_info_reg_mem_model() */
 394                 xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 395                 xa->zc_alloc->free(xa->zc_alloc, handle);
 396                 rcu_read_unlock();
 397         default:
 398                 /* Not possible, checked in xdp_rxq_info_reg_mem_model() */
 399                 break;
 400         }
 401 }
 402 
 403 void xdp_return_frame(struct xdp_frame *xdpf)
 404 {
 405         __xdp_return(xdpf->data, &xdpf->mem, false, 0);
 406 }
 407 EXPORT_SYMBOL_GPL(xdp_return_frame);
 408 
 409 void xdp_return_frame_rx_napi(struct xdp_frame *xdpf)
 410 {
 411         __xdp_return(xdpf->data, &xdpf->mem, true, 0);
 412 }
 413 EXPORT_SYMBOL_GPL(xdp_return_frame_rx_napi);
 414 
 415 void xdp_return_buff(struct xdp_buff *xdp)
 416 {
 417         __xdp_return(xdp->data, &xdp->rxq->mem, true, xdp->handle);
 418 }
 419 EXPORT_SYMBOL_GPL(xdp_return_buff);
 420 
 421 /* Only called for MEM_TYPE_PAGE_POOL see xdp.h */
 422 void __xdp_release_frame(void *data, struct xdp_mem_info *mem)
 423 {
 424         struct xdp_mem_allocator *xa;
 425         struct page *page;
 426 
 427         rcu_read_lock();
 428         xa = rhashtable_lookup(mem_id_ht, &mem->id, mem_id_rht_params);
 429         page = virt_to_head_page(data);
 430         if (xa)
 431                 page_pool_release_page(xa->page_pool, page);
 432         rcu_read_unlock();
 433 }
 434 EXPORT_SYMBOL_GPL(__xdp_release_frame);
 435 
 436 int xdp_attachment_query(struct xdp_attachment_info *info,
 437                          struct netdev_bpf *bpf)
 438 {
 439         bpf->prog_id = info->prog ? info->prog->aux->id : 0;
 440         bpf->prog_flags = info->prog ? info->flags : 0;
 441         return 0;
 442 }
 443 EXPORT_SYMBOL_GPL(xdp_attachment_query);
 444 
 445 bool xdp_attachment_flags_ok(struct xdp_attachment_info *info,
 446                              struct netdev_bpf *bpf)
 447 {
 448         if (info->prog && (bpf->flags ^ info->flags) & XDP_FLAGS_MODES) {
 449                 NL_SET_ERR_MSG(bpf->extack,
 450                                "program loaded with different flags");
 451                 return false;
 452         }
 453         return true;
 454 }
 455 EXPORT_SYMBOL_GPL(xdp_attachment_flags_ok);
 456 
 457 void xdp_attachment_setup(struct xdp_attachment_info *info,
 458                           struct netdev_bpf *bpf)
 459 {
 460         if (info->prog)
 461                 bpf_prog_put(info->prog);
 462         info->prog = bpf->prog;
 463         info->flags = bpf->flags;
 464 }
 465 EXPORT_SYMBOL_GPL(xdp_attachment_setup);
 466 
 467 struct xdp_frame *xdp_convert_zc_to_xdp_frame(struct xdp_buff *xdp)
 468 {
 469         unsigned int metasize, totsize;
 470         void *addr, *data_to_copy;
 471         struct xdp_frame *xdpf;
 472         struct page *page;
 473 
 474         /* Clone into a MEM_TYPE_PAGE_ORDER0 xdp_frame. */
 475         metasize = xdp_data_meta_unsupported(xdp) ? 0 :
 476                    xdp->data - xdp->data_meta;
 477         totsize = xdp->data_end - xdp->data + metasize;
 478 
 479         if (sizeof(*xdpf) + totsize > PAGE_SIZE)
 480                 return NULL;
 481 
 482         page = dev_alloc_page();
 483         if (!page)
 484                 return NULL;
 485 
 486         addr = page_to_virt(page);
 487         xdpf = addr;
 488         memset(xdpf, 0, sizeof(*xdpf));
 489 
 490         addr += sizeof(*xdpf);
 491         data_to_copy = metasize ? xdp->data_meta : xdp->data;
 492         memcpy(addr, data_to_copy, totsize);
 493 
 494         xdpf->data = addr + metasize;
 495         xdpf->len = totsize - metasize;
 496         xdpf->headroom = 0;
 497         xdpf->metasize = metasize;
 498         xdpf->mem.type = MEM_TYPE_PAGE_ORDER0;
 499 
 500         xdp_return_buff(xdp);
 501         return xdpf;
 502 }
 503 EXPORT_SYMBOL_GPL(xdp_convert_zc_to_xdp_frame);